You could ingest them with AWS Textract and have predictability and formatting in the format of your choice. Using LLMs for this is lazy and generates unpredictable and non-deterministic results.